一、张量数据类型
1.1 pytorch与python数据类型对比
python | pytorch |
---|---|
Int | IntTensor of size() |
float | FloatTensor of size() |
Int array | IntTensor of size [d1,d2,…] |
Float array | FloatTensor of size [d1,d2,…] |
string | ont-hot or Embedding(Word2Vec,glove) |
由于PyTorch不是一个完备的语言库,它是面向数据计算的GPU加速库,因此没有内建对string的支持,在PyTorch中可用以下两种方式来表示string:
- One-hot
- Embedding
- Word2vec
- glove
1.2 pytorch内建的数据类型
Data tyoe | dtype | CPU tensor | GPU tensor |
---|---|---|---|
32-bit floating point | torch.float32 or torch.float | torch.FloatTensor |
torch.cuda.FloatTensor |
64-bit floating point | torch.float64 or torch.double | torch.DoubleTensor |
torch.cuda.DoubleTensor |
16-bit floating point | torch.float16 or torch.half | N/A | torch.cuda.HalfTensor |
8-bit integer (unsigned) | torch.uint8 | torch.ByteTensor |
torch.cuda.ByteTensor |
8-bit integer (signed) | torch.int8 | torch.CharTensor |
torch.cuda.CharTensor |
16-bit integer (signed) | torch.int16 or torch.short | torch.ShortTensor |
torch.cuda.ShortTensor |
32-bit integer (signed) | torch.int32 or torch.int | torch.IntTensor |
torch.cuda.IntTensor |
64-bit integer (signed) | torch.int64 or torch.long | torch.LongTensor |
torch.cuda.LongTensor |
Tpye check
- Tensor.type()
- type(Tensor)
- isinstance(Tensor,torch.FloatTensor)
>>> a = torch.randn(2,3) >>> a.type() 'torch.FloatTensor' >>> type(a) <class 'torch.Tensor'> >>> isinstance(a,torch.FloatTensor) True >>> isinstance(a,torch.cuda.DoubleTensor) False >>> data = a.cuda() #x.cuda()返回一个gpu的引用 >>> isinstance(data,torch.cuda.DoubleTensor) >>> True
标量:Dimension 0/rank 0,Deep Learning中常用于loss表示
>>> torch.tensor(2.2)#.tensor接受的是数据的内容 tensor(2.2000) >>> a = torch.tensor(2.2) >>> a.shape torch.Size([]) >>> len(a.shape) 0 >>> a.dim() 0 >>> a.size() torch.Size([])
Dimension 1/rank 1,Deep Learning中常用于Bias表示
>>> torch.tensor([1.1]) tensor([1.1000]) >>> torch.tensor([1.1,2.2]) tensor([1.1000, 2.2000]) >>> torch.FloatTensor(2) #接受的是size,random初始化 tensor([2.2758e-07, 4.5682e-41]) >>> data = np.ones(2) >>> data array([1., 1.]) >>> torch.from_numpy(data) tensor([1., 1.], dtype=torch.float64)
Dimension 2
>>> a = torch.randn(2,3) #FloatTensor(2,3) >>> a tensor([[-0.1680, 0.4534, -0.4045], [-1.0437, -0.4634, 0.7432]]) >>> a.shape torch.Size([2, 3]) >>> a.size(0) 2 >>> a.size(1) 3 >>> a.shape[0] 2 >>> a.shape[1] 3
Dim 3:RNN Input with Batch
>>> a=torch.rand(1,2,3) >>> a tensor([[[0.2226, 0.0342, 0.1301], [0.6371, 0.6930, 0.9356]]]) >>> a.shape torch.Size([1, 2, 3]) >>> a[0] #[2,3] tensor([[0.2226, 0.0342, 0.1301], [0.6371, 0.6930, 0.9356]]) >>> list(a.shape) [1, 2, 3]
Dim4:CNN[b,c,h,w]
numbel是指tensor占用内存的数量
>>> a = torch.rand(2,3,28,28) >>> a tensor([[[[0.7190, 0.8762, 0.3667, ..., 0.8682, 0.5834, 0.7012], [0.4110, 0.5633, 0.1516, ..., 0.6877, 0.1930, 0.9480], [0.0063, 0.8593, 0.4722, ..., 0.4012, 0.8891, 0.0254], ..., ..., [0.3267, 0.8081, 0.5329, ..., 0.3658, 0.9325, 0.6759], [0.4113, 0.8107, 0.9934, ..., 0.2609, 0.1763, 0.5233], [0.7673, 0.3748, 0.0287, ..., 0.0348, 0.0529, 0.8054]]]]) >>> >>> a.shape torch.Size([2, 3, 28, 28]) >>> a.numel() 4704 >>> a.dim() #len(a.shape) 4 >>> a=torch.tensor(1) >>> a.dim() 0
创建Tensor
- Import form
>>> a=np.array([2,3,3]) >>> torch.from_numpy(a) tensor([2, 3, 3]) >>> a=np.ones([2,3,3]) >>> torch.from_numpy(a) tensor([[[1., 1., 1.], [1., 1., 1.], [1., 1., 1.]], [[1., 1., 1.], [1., 1., 1.], [1., 1., 1.]]], dtype=torch.float64)
- Import from List
torch.tensor([2.,3.2]) tensor([2.0000, 3.2000]) torch.FloatTensor([2.,3.2]) #这种方式不推荐使用,接受具体的数据推荐使用torch,tensor tensor([2.0000, 3.2000]) torch.Tensor([2.,3.2]) tensor([2.0000, 3.2000]) torch.tensor([[2.,3.2],[1.,22.3]]) tensor([[ 2.0000, 3.2000], [ 1.0000, 22.3000]])
torch.FloatTensor和torch.Tensor基本是一样的,都是接受书的维度shape创建Tensor,而小写的torch.tensor是接受具体的数据
uninitialized
- Torch.empty(2,3)
- Torch.FloatTensor(d1,d2,d3)
- Torch.IntTensor(d1,d2,d3)
注意未初始化的api在使用的时候,若后续没有赋值操作覆盖,将会使得随机初始化的值变的非常大或者非常小(troch.nan/inf)
随机初始化
- torch.rand
- torch.rand_like
- torch.randint
- torch.randn:正态分布
- torch.normal:自定义分布
>>> torch.rand(3,3) tensor([[0.3628, 0.4771, 0.5067], [0.6593, 0.6323, 0.9157], [0.5882, 0.6289, 0.4311]]) >>> a=torch.rand(3,3) >>> torch.rand_like(a) tensor([[0.5168, 0.9998, 0.1509], [0.6104, 0.5265, 0.7917], [0.3084, 0.9142, 0.0541]]) >>> torch.randint(1,10,[3,3]) tensor([[2, 2, 6], [3, 9, 1], [4, 5, 4]]) >>> torch.normal(mean=torch.full([10],0),std=torch.arange(1,0,-0.1)) tensor([-2.1002, -0.2133, 0.9746, 0.6781, 0.3725, 0.6669, 0.4720, 0.7872, 0.0643, 0.0143])
set default type
>>> torch.tensor([1.2,3]).type() 'torch.FloatTensor' #pytorch默认的数据类型是FloatTensor >>> torch.set_default_tensor_type(torch.DoubleTensor) >>> torch.tensor([1.2,3]).type() 'torch.DoubleTensor' #增强学习一般使用double
二、torch常用函数
- Import form
在使用Tensor时,我们首先要掌握如何使用Tensor来定义不同数据类型的变量。和Numpy差不多,PyTorch中的Tensor也有自己的数据类型定义方式,常用的如下:
1.张量Tensors
- torch.is_tensor(obj) #判断是否为张量,如果是pytorch张量,则返回True
- torch.is_storage(obj) #判断是否为pytorch Storage,如何是,则返回True
- torch.set_default_tensor_type(t)
- torch.numel(input)->int
- torch.set_printoptions(precision=None, threshold=None, edgeitems=None, linewidth=None, profile=None) #设置打印选项。 完全参考自Numpy。
2.创建操作 Creation Ops
torch.eye
torch.eye(n, m=None, out=None)->Tensor
返回一个2维张量(单位矩阵),对角线数字为1,其它位置为0
参数说明:
n
(int) – 行数m
(int, 可选) – 列数.如果为None,则默认为nout
(Tensor,可选) - 输出张量
>>> torch.eye(4)
tensor([[1., 0., 0., 0.],
[0., 1., 0., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.]])
torch.zero
torch.zeros(*sizes, out=None) → Tensor
返回一个全0的张量,形状由可变参数sizes
定义。
>>> torch.zeros(2,3)
tensor([[0., 0., 0.],
[0., 0., 0.]])
>>> torch.zeros(5)
tensor([0., 0., 0., 0., 0.])
torch.ones
torch.ones(*sizes, out=None) → Tensor #返回一个全为1的张量,形状由可变参数
sizes
定义。
>>> torch.ones(2,3)
tensor([[1., 1., 1.],
[1., 1., 1.]])
>>> torch.ones(5)
tensor([1., 1., 1., 1., 1.])
torch.from_numpy
torch.from_numpy(ndarray) → Tensor
将numpy.ndarray
转换为Tensor
。 返回的张量tensor和numpy的ndarray共享同一内存空间
。修改一个会导致另外一个也被修改。返回的张量不能调整大小
。
>>> a= np.array([1,2,4])
>>> t = torch.from_numpy(a)
>>> t
tensor([1, 2, 4])
>>> t[1]=10
>>> a
array([ 1, 10, 4])
torch.linspace
torch.linspace(start, end, steps=100, out=None) → Tensor #返回start和end之间长度为
steps
的一维张量
参数说明:
start (float)
– 点集的起始值end (float)
– 点集的最终值steps (int)
– 在start
和end
间的采样数,即返回多少个数out (Tensor, 可选的)
– 结果张量
>>> torch.linspace(1.0,10,steps=5)
tensor([ 1.0000, 3.2500, 5.5000, 7.7500, 10.0000])
>>> torch.linspace(-10,10,steps=5,out=None)
tensor([-10., -5., 0., 5., 10.])
torch.logspace
torch.logspace(start, end, steps=100, out=None) → Tensor
返回一个1维张量,包含在区间\(10^{start}\) 和\(10^{end}\)上以对数刻度均匀间隔的steps
个点。 输出1维张量的长度为steps
,参数说明同torch.linspace
>>> torch.logspace(start=-10,end=10,steps=5)
tensor([1.0000e-10, 1.0000e-05, 1.0000e+00, 1.0000e+05, 1.0000e+10])
>>> torch.logspace(start=0.1,end=1.0,steps=5)
tensor([ 1.2589, 2.1135, 3.5481, 5.9566, 10.0000])
torch.FloatTensor
此变量用于生成数据类型为浮点型的Tensor,传递给torch.FloatTensor的参数可以是一个列表,也可以是一个维度值。torch.IntTensor与此类似。
>>> import torch
>>> a = torch.FloatTensor(2,3)
>>> print(a)
-0.1171 0.0000 -0.1171
0.0000 0.0000 0.0000
[torch.FloatTensor of size 2x3]
>>> b = torch.FloatTensor([2,3,4,5])
>>> print(b)
2
3
4
5
[torch.FloatTensor of size 4]
torch.rand
用于生成数据类型为
浮点型且维度指定
的随机Tensor,和在Numpy中使用numpy.rand生成随机数的方法类似,随机生成的浮点数据在0~1区间均匀分布
。
torch.rand(*sizes, out=None) → Tensor
返回一个张量,填充在[0,1]区间的一组均匀分布随机数。 Tensor的形状由变量
sizes
定义。
>>> torch.rand(4)
tensor([0.0662, 0.7079, 0.4197, 0.2257])
>>> torch.rand(2,3,out=None)
tensor([[0.8174, 0.8959, 0.2713],
[0.5343, 0.0275, 0.7712]])
>>> torch.rand(2,3)
torch.randn
torch.randn(*sizes, out=None) → Tensor
用于生成数据类型为浮点型且维度指定
的随机Tensor,和在Numpy中使用numpy.randn生成随机数的方法类似,随机生成的浮点数的取值满足均值为0,方差为1的正态分布
。
>>> torch.randn(4)
tensor([-1.4524, 0.9949, -1.4038, 0.8059])
>>> torch.randn(2,3)
tensor([[ 0.9186, 1.0356, 1.1103],
[-2.0057, -0.9032, 0.6453]])
torch.randperm
torch.randperm(n, out=None) → LongTensor
输入参数n(int) – 上限(独占),即最大值
,返回一个从0
到n -1
的随机整数排列,随机打散。
random.shuffle
>>> torch.randperm(4)
tensor([3, 1, 2, 0])
>>> a=torch.rand(2,3)
>>> b=torch.rand(2,2)
>>> idx=torch.randperm(2)
>>> idx
tensor([1, 0])
>>> idx
tensor([1, 0])
>>> a,b
(tensor([[0.8094, 0.8389, 0.5666],
[0.6812, 0.5959, 0.4951]]), tensor([[0.4863, 0.5345],
[0.1246, 0.2468]]))
>>> a[idx]
tensor([[0.6812, 0.5959, 0.4951],
[0.8094, 0.8389, 0.5666]])
>>> b[idx]
tensor([[0.1246, 0.2468],
[0.4863, 0.5345]])
torch.arange
torch.arange(start, end, step=1, out=None) → Tensor([start;end))
返回一个1维张量,长度为floor((end?start)/step),floor代表向下取整。包含从start
到end
,以step
为步长的一组序列值(默认步长为1)。
>>> torch.arange(1,4)
tensor([1, 2, 3])
>>> torch.arange(1,2.5,0.5)
tensor([1.0000, 1.5000, 2.0000])
torch.range
torch.range(start, end, step=1, out=None) → Tensor([start;end]),现在torch.range将要废弃,
返回一个1维张量,长度为floor((end?start)/step)+1,其中floor代表向下取整数。从start
开始,end
为结尾,以step
为步长的一组值。 step
是两个值之间的间隔,即 Xi+1=Xi+step
>>> import torch
>>> a = torch.range(2,8,1)
__main__:1: UserWarning: torch.range is deprecated in favor of torch.arange and will be removed in 0.5. Note that arange generates values in [start; end), not [start; end].
>>> print(a)
tensor([2., 3., 4., 5., 6., 7., 8.])
torch.full
>>> torch.full([2,3],7)
tensor([[7., 7., 7.],
[7., 7., 7.]])
>>> torch.full([],7) #给标量赋值
tensor(7.)
>>> torch.full([1],7)
tensor([7.])
>>> torch.full([2],7)
tensor([7., 7.])
3.Indexing, Slicing, Joining, Mutating Ops
- dim 0 first
- select first/last N
- select by steps
- select by specific index
- ...
dim 0 first
>>> a = torch.rand(4,3,28,28)
>>> a[0].shape
torch.Size([3, 28, 28])
>>> a[0,0].shape
torch.Size([28, 28])
>>> a[0,0,2,4]
tensor(0.2409)
select first/last N
>>> a[:2].shape
torch.Size([2, 3, 28, 28])
>>> a[:2,:1,:,:].shape
torch.Size([2, 1, 28, 28])
>>> a[:2,1:,:,:].shape
torch.Size([2, 2, 28, 28])
>>> a[:2,-1:,:,:].shape
torch.Size([2, 1, 28, 28])
select by steps
>>> a[:,:,0:28:2,:].shape
torch.Size([4, 3, 14, 28])
>>> a[:,:,0:28:2,0:14].shape
torch.Size([4, 3, 14, 14])
>>> a[:,:,0:28:2,2:14].shape
torch.Size([4, 3, 14, 12])
通用形式:start : end : step
select by specific index
torch.index_select
torch.index_select(input, dim, index, out=None) → Tensor
参数:
- input (Tensor) – 输入张量
- dim (int) – 索引的轴
- index (LongTensor) – 包含索引下标的一维张量
- out (Tensor, optional) – 目标张量
沿着指定维度对输入进行切片,取index
中指定的相应项(index
为一个LongTensor),然后返回到一个新的张量, 返回的张量与原始张量_Tensor_有相同的维度(在指定轴上)。
注意: 返回的张量不与原始张量共享内存空间。
>>> x=torch.randn(3,4)
>>> x
tensor([[ 0.4878, -0.7477, -0.2496, 1.1835],
[-1.8953, 0.3530, 0.1122, 0.9137],
[ 0.4686, 0.5230, -1.1191, -1.0911]])
>>> indices = torch.LongTensor([0,2])
>>> indices
tensor([0, 2])
>>> torch.index_select(x,1,indices)
tensor([[ 0.4878, -0.2496],
[-1.8953, 0.1122],
[ 0.4686, -1.1191]])
>>> torch.index_select(x,1,indices)
tensor([[ 0.4878, -0.2496],
[-1.8953, 0.1122],
[ 0.4686, -1.1191]])
>>> torch.index_select(x,0,indices)
tensor([[ 0.4878, -0.7477, -0.2496, 1.1835],
[ 0.4686, 0.5230, -1.1191, -1.0911]])
...
>>> a.shape
torch.Size([4, 3, 28, 28])
>>> a[...].shape
torch.Size([4, 3, 28, 28])
>>> a[0,...].shape #等同于a[0]
torch.Size([3, 28, 28])
>>> a[:,1,...].shape
torch.Size([4, 28, 28])
>>> a[0,...,::2].shape #等同于a[0,:,:,::2]
torch.Size([3, 28, 14])
>>> a[...,:2].shape
torch.Size([4, 3, 28, 2])
select by mask
torch.masked_select
torch.masked_select(input, mask, out=None) → Tensor
- input (Tensor) – 输入张量
- mask (ByteTensor) – 掩码张量,包含了二元索引值
- out (Tensor, optional) – 目标张量
根据掩码张量mask
中的二元值,取输入张量中的指定项( mask
为一个 ByteTensor),将取值返回到一个新的1D张量,
张量 mask
须跟input
张量有相同数量的元素数目,但形状或维度不需要相同。 注意: 返回的张量不与原始张量共享内存空间。
>>> x = torch.randn(3,4)
>>> x
tensor([[ 0.3132, -0.4580, -0.3642, 0.7394],
[-0.2821, 1.9086, -0.9687, 1.6009],
[ 0.9800, -0.8546, -0.8855, -0.3807]])
>>> mask = x.ge(0.5)
>>> mask
tensor([[False, False, False, True],
[False, True, False, True],
[ True, False, False, False]])
>>> torch.masked_select(x,mask)
tensor([0.7394, 1.9086, 1.6009, 0.9800])
>>> torch.masked_select(x,mask).shape
torch.Size([4])
select by flatten index
>>> src = torch.tensor([[4,3,5],[6,7,8]])
>>> torch.take(src,torch.tensor([0,2,5]))
tensor([4, 5, 8])
torch.nonzero
torch.nonzero(input, out=None) → LongTensor
参数:
- input (Tensor) – 源张量
- out (LongTensor, optional) – 包含索引值的结果张量
返回一个包含输入input
中非零元素索引的张量。输出张量中的每行包含输入中非零元素的索引。如果输入input
有n
维,则输出的索引张量output
的形状为 z x n, 这里 z 是输入张量input
中所有非零元素的个数。
>>> torch.nonzero(torch.Tensor([1,1,1,0,1])) #返回一维张量中非零的索引值
tensor([[0],
[1],
[2],
[4]])
#输入张量input有2(n)维,非零元素有4(z)个,所以输出张量output的shape为4× 2
>>> torch.nonzero(torch.Tensor([[0.6,0.0,0.0,0.0],[0.0,0.4,0.0,0.0],[0.0,0.0,1.2,0.0],[0.0,0.0,5,0.0]]))
tensor([[0, 0],
[1, 1],
[2, 2],
[3, 2]]) #表明input的第4行的index2的为非零数据
维度变换
- view/reshape(view保证numel不变即可)
- Squeeze/unsqueeze(删减/增加)
- Transpose/t/permute
- Expand/repeat
torch.view/torch.reshape
view操作需保证数据有实际意义,不然会将数据造成破坏,数据的存储、维度顺序非常重要
>>> a=torch.rand(4,1,28,28) #4张灰度图片
>>> a.view(4,28*28) #(4,1*28*28)将channel和像素值合并在一起=》[4,784]适合于全连接层
tensor([[0.2301, 0.9408, 0.3547, ..., 0.9387, 0.0988, 0.9476],
[0.2724, 0.6440, 0.0037, ..., 0.7575, 0.1136, 0.7190],
[0.6347, 0.9259, 0.4316, ..., 0.8456, 0.2670, 0.6662],
[0.0801, 0.3157, 0.4126, ..., 0.4852, 0.2193, 0.8381]])
>>> a.view(4,28*28).shape
torch.Size([4, 784])
>>> a.view(4*28,28).shape
torch.Size([112, 28])
>>> a.view(4*1,28,28).shape
torch.Size([4, 28, 28]) #(4*1,28*28),指关注featuremap
>>> b=a.view(4,784)
>>> b.view(4,28,28,1) #logic Bug,将维度信息丢失掉了
tensor([[[[0.2301],
[0.9408],
[0.3547],
...,
[0.4852],
[0.2193],
[0.8381]]]])
torch.unsqueeze
torch.unsqueeze(input, dim, out=None)
参数:
- tensor (Tensor) – 输入张量
- dim (int) – 插入维度的索引
- out (Tensor, optional) – 结果张量
返回一个新的张量,对输入的指定位置插入维度。注意: 返回张量与输入张量共享内存,所以改变其中一个的内容会改变另一个。如果dim
为负,则将会被转化dim+input.dim()+1,指定位置在 1[-a.dim()-1,a.dim()+1]=>[-5,5]之间变化。
>>> a.shape
torch.Size([4, 1, 28, 28])
>>> a.unsqueeze(0).shape
torch.Size([1, 4, 1, 28, 28])
>>> a.unsqueeze(-1).shape
torch.Size([4, 1, 28, 28, 1])
>>> a.unsqueeze(-4).shape
torch.Size([4, 1, 1, 28, 28])
>>>
>>> a.unsqueeze(-5).shape
torch.Size([1, 4, 1, 28, 28])
>>> a.unsqueeze(5).shape
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
IndexError: Dimension out of range (expected to be in range of [-5, 4], but got 5)
>>> a=torch.tensor([1.2,2.3])
>>> a.unsqueeze(-1) # [2,1]
tensor([[1.2000],
[2.3000]])
>>> a.unsqueeze(0) #[1,2]
tensor([[1.2000, 2.3000]])
example:
>>> b=torch.rand(32) #bias相当于给每个channel上的所有像素增加一个偏置
>>> f=torch.rand(4,32,14,14)
>>> b=b.unsqueeze(1).unsqueeze(2).unsqueeze(0)
>>> b.shape
torch.Size([1, 32, 1, 1])
torch.squeeze
>>> b.shape
torch.Size([1, 32, 1, 1])
>>> b.squeeze().shape
torch.Size([32])
>>> b.squeeze(0).shape
torch.Size([32, 1, 1])
>>> b.shape
torch.Size([1, 32, 1, 1])
>>> b.squeeze(-1).shape
torch.Size([1, 32, 1])
>>> b.squeeze(1).shape #32是不能被挤压的,只能挤压1的维度
torch.Size([1, 32, 1, 1])
>>> b.squeeze(-4).shape
torch.Size([32, 1, 1])
torch.unbind
torch.unbind(tensor, dim=0)[source]
参数:
- tensor (Tensor) – 输入张量
- dim (int) – 删除的维度
移除指定维后,返回一个元组,包含了沿着指定维切片后的各个切片
torch.expand/torch.repeat
- Expand:broadcasting【不会主动复制数据,只在需要的时候,这是推荐的一种方式;进行操作的时候需要前后dim一致,只能从dim1->dimN;-1表示dim不变】
- Repeat:memory copied(给出的参数p为在相应的维度上拷贝p次)
###expand
>>> a=torch.rand(4,32,14,14)
>>> b.shape
torch.Size([1, 32, 1, 1])
>>> b.expand(4,32,14,14).shape
torch.Size([4, 32, 14, 14])
>>> b.expand(-1,32,-1,-1).shape
torch.Size([1, 32, 1, 1])
>>> b.expand(-1,32,-1,-4).shape #-4是bug,在最新版的facebook已经被修复了
torch.Size([1, 32, 1, -4])
##repeat
>>> b.shape
torch.Size([1, 32, 1, 1])
>>> b.repeat(4,32,1,1).shape
torch.Size([4, 1024, 1, 1])
>>> b.repeat(4,1,1,1).shape
torch.Size([4, 32, 1, 1])
>>> b.repeat(4,1,32,32).shape
torch.Size([4, 32, 32, 32])
torch.t
torch.t(input, out=None) → Tensor
参数:
- input (Tensor) – 输入张量
- out (Tensor, optional) – 结果张量
输入一个矩阵(2维张量),并转置0, 1维。 可以被视为函数transpose(input, 0, 1)
的简写函数,但是torch.t只适用于2D。
torch.transpose
torch.transpose(input, dim0, dim1, out=None) → Tensor
- input (Tensor) – 输入张量
- dim0 (int) – 转置的第一维
- dim1 (int) – 转置的第二维
返回输入矩阵input
的转置。交换维度dim0
和dim1
。 输出张量与输入张量共享内存,所以改变其中一个会导致另外一个也被修改。高维矩阵的转置常用torch.transpose
>>> a=torch.rand(4,3,32,32)
>>> a1.shape
torch.Size([4, 32, 32, 3])
>>> a1 = a.transpose(1,3).view(4,3*32*32).view(4,3,32,32)#[b,c,h,w]->[b,w,h,c]->[b,c,w,h]维度信息与原来的存储信息不一致,报错
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.
>>> a1 = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,3,32,32) #contiguous将数据变为连续,同上,维度污染,错误
>>> a2 = a.transpose(1,3).contiguous().view(4,3*32*32).view(4,32,32,3).transpose(1,3)
>>> a1.shape,a2.shape
(torch.Size([4, 3, 32, 32]), torch.Size([4, 3, 32, 32]))
>>> torch.all(torch.eq(a,a1))
tensor(False)
>>> torch.all(torch.eq(a,a2))
tensor(True)
permute
permute()参数信息是原来Tensor的维度信息
[b,h,w,c]是numpy存储图片的格式,只有这个格式才能导出numpy
>>> a=torch.rand(4,3,28,28) #[b,c,h,w]
>>> a.transpose(1,3).shape
torch.Size([4, 28, 28, 3])
>>> b=torch.rand(4,3,28,32)
>>> b.transpose(1,3).shape #[b,c,h,w]->[b,w,h,c]
torch.Size([4, 32, 28, 3])
>>> b.transpose(1,3).shape #[b,w,h,c]->[b,c,h,w]
torch.Size([4, 32, 28, 3])
>>> b.permute(0,2,3,1).shape #[b,c,h,w]->[b,h,w,c]
torch.Size([4, 28, 32, 3])
原文地址:https://www.cnblogs.com/lyszyl/p/12161819.html